Originally presented at:
OCEANS 2018 MTS/IEEE Charleston, 22-25 October 2018,
doi: 10.1109/OCEANS.2018.8604521
SCIENCE AND TECHNOLOGY ORGANIZATION
CENTRE FOR MARITIME RESEARCH AND EXPERIMENTATION
Reprint Series CMRE-PR-2019-025
On adaptive modulation for low SNR underwater acoustic communications
Konstantinos Pelekanakis, Luca Cazzanti
May 2019
About CMRE
The Centre for Maritime Research and Experimentation (CMRE) is a world-class NATO scientific research and experimentation facility located in La Spezia, Italy.
The CMRE was established by the North Atlantic Council on 1 July 2012 as part of the NATO Science & Technology Organization. The CMRE and its predecessors have served NATO for over 50 years as the SACLANT Anti-Submarine Warfare Centre, SACLANT Undersea Research Centre, NATO Undersea Research Centre (NURC) and now as part of the Science & Technology Organization.
CMRE conducts state-of-the-art scientific research and experimentation ranging from concept development to prototype demonstration in an operational environment and has produced leaders in ocean science, modelling and simulation, acoustics and other disciplines, as well as producing critical results and understanding that have been built into the operational concepts of NATO and the nations.
CMRE conducts hands-on scientific and engineering research for the direct benefit of its NATO Customers. It operates two research vessels that enable science and technology solutions to be explored and exploited at sea. The largest of these vessels, the NRV Alliance, is a global class vessel that is acoustically extremely quiet.
CMRE is a leading example of enabling nations to work more effectively and efficiently together by prioritizing national needs, focusing on research and technology challenges, both in and out of the maritime environment, through the collective Power of its world-class scientists, engineers, and specialized laboratories in collaboration with the many partners in and out of the scientific domain.
Copyright © IEEE, 2018. NATO member nations have unlimited rights to use, modify, reproduce, release, perform, display or disclose these materials, and to authorize others to do so for government purposes. Any reproductions marked with this legend must also reproduce these markings. All other rights and uses except those permitted by copyright law are reserved by the copyright owner.
NOTE: The CMRE Reprint series reprints papers and articles published by CMRE authors in the open literature as an effort to widely disseminate CMRE products. Users are encouraged to cite the original article where possible.
On Adaptive Modulation for low SNR UnderwaterAcoustic Communications
Konstantinos Pelekanakis1 and Luca Cazzanti2
1 NATO STO Centre for Maritime Research and Experimentation (CMRE), La Spezia 19126, Italy
e-mail: [email protected]
2 Amplero, Seattle, WA 98101, USA
e-mail: [email protected]
Abstract—This paper deals with adaptive underwater acoustic(UWA) communications where the receiver must operate at lowsignal-to-noise ratios (SNRs). The proposed modem is equippedwith a set of direct sequence spread spectrum (DSSS) signals ofvarious coding rates and modulation orders. A channel-estimate-based decision feedback equalizer (CEB-DFE) is used at thereceiver. We address the challenge of achieving high spectralefficiencies subject to a combination of bit-error rate (BER) andSNR constraints. To this end, adaptive selection of signals isachieved based on their BER prediction via boosted trees. Thisensemble of trees learns directly from the received data andrelates the BER with signal characteristics and channel metrics.The efficiency of the boosted trees is validated by post-processingthousands of acoustic signals recorded in the Gulf of La Spezia,Italy. 10-20 times faster communications as compared to a modemwith a fixed rate transmission is achieved.
Index Terms—Adaptive Modulation and Coding (AMC), ma-chine learning, regression trees, Low Probability of Intercept(LPI).
I. INTRODUCTION
Underwater acoustic networks (UANs) typically cover a
large geographic area and use acoustic signals to link the
nodes together. The open nature of the acoustic channel
allows unauthorized nodes to capture data or disrupt network
functionality and therefore protection against security attacks
is of paramount importance. Eavesdropping is a common
type of security attack where an adversary uses a device that
does not belong to the network to access to the contents
of the communications message. The typical approach is to
communicate at very low signal-to-noise ratio (SNR) by means
of direct sequence spread spectrum (DSSS) modulation [1].
In this modulation technique, the information symbols are
multiplied/spread by a code sequence resulting in a wideband
transmitted signal. At the receiver side, the code sequence
is used as a matched filter. Since the probability to detect
a signal is SNR-dependent and the spreading code is known
only to legitimate receivers, low probability of intercept (LPI)
communication is said to be achieved with DSSS modulation
at low transmit power.
A major challenge in applying DSSS in UWA communi-
cations is synchronization in the presence of long and time-
varying reverberation. In [2], the authors have applied time-
updated passive-phase conjugation (PPC) achieving a bit-error
rate (BER) of 10−2 at -12 dB SNR. The same authors, have
applied incoherent detection of DSSS signals to achieve good
communications at -10 dB [3]. In [4], the authors have tested
spreading codes suitable for a coherent RAKE receiver. The
reported performance is 5 · 10−3 BER at -8 dB SNR. In [5],
the authors have combined a Reed-Solomon code with M-ary
orthogonal code keying (M-OCK) to achieve 35.63 bps at -
14 dB SNR. In [6] a novel way to improve the LPI capability
of communications based on chaotic sequences has been
demonstrated. It is worthy to note that any coding technique
that trades power for bandwidth can be used to enhance the
LPI capability of the waveform. For example, the researchers
in [7] have tested a multi-band equaliser system running on
1/3-rate turbo codes. The reported performance is 75 bps at
-12 dB SNR with a single hydrophone.
An urgent requirement of LPI communications is the smart
selection of the waveform length and bit rate that best fits the
application requirements as well as the channel conditions.
In other words, the tradeoff between throughput, reliability
and target SNR needs to be quantified. Given the vast variety
of underwater acoustic channels, closed-form formulas that
predict the BER of a specific modulation technique based
on the receiver characteristics are intractable in practice.
Machine learning techniques that are purely data driven and
capture the nonlinear effects (environmental and hardware-
related) on the modem performance provide a promising
avenue for developing adaptive modulation strategies. Some
indicative machine learning techniques that have been applied
for adaptive modulation and coding (AMC) in wireless radio
are: neural networks [8], [9], support vector machines [10],
decision trees classifiers [11] and kernel regression [12]. In the
underwater acoustic domain, applications of machine learning
algorithms are very scarce. A Bayesian inference algorithm has
been applied in [13]. The same authors in [14] framed AMC
as a multi-armed bandit problem to address the explorationvs exploitation dilemma. However, in both studies the channel
state information available at the receiver was not considered
to improve link adaptation.
Building upon our previous work presented in [15], we
design a single-carrier modem equipped with seven DSSS
signals of various bit rates. The goal is to transmit the signal
with the maximal bit rate in the next transmission slot based on
user-defined SNR and BER constraints. We use boosted trees
CMRE Reprint Series CMRE-PR-2019-025
1
to learn the relationship between the measured BER with the
signal parameters, the received SNR, and other related channel
metrics which characterize the signal distortion. To validate
our approach, we apply the boosted tree on a large amount
of recorded signals transmitted during the Littoral Acoustic
Communications Experiment 2017 (LACE17).
The paper is organised as follows. In Section II, the modem
design and the experimental setup of LACE17 are presented.
The proposed channel metrics that aid boosted trees to perform
BER prediction are discussed in Section III. The BER results
as well as the BER prediction capability of boosted trees
are shown in Section IV. Finally, the paper is concluded in
Section V.
II. MODEM DESIGN AND EXPERIMENTAL SETUP
We design a modem that is a able to transmit and receive
waveforms of different spectral efficiencies as a result of
different coded modulation schemes, spreading sequences and
baud rates. The transmitter block diagram can be seen in
Figure 1(a). The information-bearing bit sequence is encoded
and mapped into M-ary Phase-shift keying (PSK) channel
symbols. For 2-PSK symbols, a 1/2-rate convolutional encoder
is used. For 4- and 8-PSK symbols, Trellis Coded Modulation
(TCM) with rate 1/2 and 2/3, respectively, is used. The PSK
sequence is interleaved (shuffled) and then each PSK symbol
is multiplied by a Kasami sequence of values -1 and +1. The
resulting sequence {d(k)} is pulse-shaped via a raised cosine
filter with chip interval T and roll-off factor α. The baseband
waveform is simultaneously modulated onto M carriers and
transmitted in the water. The passband transmitted signal is
given by
u(t) =∑k
d(k)g(t− kT ), (1)
x(t) = Re
{M∑
m=1
u(t)ej2πfmt
}, (2)
where g(t) is the raised cosine pulse, u(t) is the baseband
signal and fm denotes the carrier frequency for the mth sub-
band. Note the mth passband signal occupies the frequency
range fm± (1+α)/(2T ) . The carrier frequencies are chosen
as follows [7]
fm = fc + (m− M + 1
2)B, (3)
where B is the passband operational bandwidth and fc is
its middle frequency. Table I summarises the signals used
in the modem. The signals BPSK-1, BPSK-2, BPSK-3, and
BPSK-4 use 1/2-rate convolutional codes combined with 2-
PSK. The signals QPSK-1 and QPSK-2 use 1/2-rate TCM
codes combined with 4-PSK. The signal 8PSK-1 uses 2/3-rate
TCM code combined with 8-PSK.
The block diagram of the receiver is shown in Figure 1(b).
Initial processing involves shifting the acquired signal to
baseband and coarse synchronization. Inter-chip interference
due to time-varying multipath is tackled by our channel-
estimate-based decision feedback equalizer (CEB-DFE). The
TABLE ILIST OF SIGNALS.
Signal nameKasami
length
Baud rate
(symbols/s)
Carrier
(kHz)
Bit rate
(bps)
BPSK-1 15 3500 10.4 116
BPSK-2 63 3500 10.4 27
BPSK-3 15 1750 9.26, 11.53 58
BPSK-4 63 1750 9.26, 11.53 13
QPSK-1 15 3500 10.4 233
QPSK-2 63 3500 10.4 55
8PSK-1 15 3500 10.4 466
CEB-DFE performs three sequential stages: (1) mean time
scale compensation; (2) sparse channel estimation; (3) decision
feedback equalization. Adaptation of these three processing
stages is performed with a single estimation error. This error
is the difference between the estimated M-ary PSK symbol
(or chip, since the DFE operates on the spread sequence)
before mapping and its value after mapping onto the signal
constellation. At the beginning of each signal reception, the
CEB-DFE runs in training mode and the true PSK symbols
are known. After the training period, the on line decisions
of the CEB-DFE are used to estimate the error. Channel
estimation is performed in each sub-band independently via
the Improved-Proportionate M-estimate Affine Projection Al-
gorithm (IPMAPA). Joint sub-band equalization is performed
based on the Recursive Least Squares (RLS) algorithm. The
interested reader is directed to [16] for additional information.
The soft symbol output of the CEB-DFE is permuted back
(deinterleaving) to the original order before spreading. After
de-interleaving the received sequence is matched filtered with
the spreading code (de-spreading). The de-spread output is fed
into the soft Viterbi decoder to compute the BER.
The BER performance of the modem was tested with
acoustic data recorded from LACE17. The LACE17 trials
took place between November 21st and November 25th in the
Gulf of La Spezia, Italy. A map of the experimental area is
shown in Figure 2. The depth of the area was about 10-12 m.
Two sources were mounted on two rigid tripods 2 m above
the seabed. The operational bandwidth of each source was
about 8-12 kHz. In addition to the sources, four hydrophones
were attached on each tripod at 50 cm, 1.1 m, 1.7 m and
2.3 m above the seabed. A third tripod with one hydrophone
located 1.4 m above the seabed was also deployed. To test
different ranges, one tripod with a source was deployed at
different positions within the red area of Figure 2. As a result,
the point-to-point links varied between 200 m and 800 m.
It is important to mention two phenomena that took place
during LACE17. The first is the occurrence of a storm with
strong winds (about 5 m/s) during the 25th, which led to
choppy seas. The second phenomenon is that the ambient
noise was not stationary because of shipping and construction
CMRE Reprint Series CMRE-PR-2019-025
2
�
�
Fig. 1. (a) Transmitter block diagram. (b) Receiver block diagram.
1.2 km
Fig. 2. LACE17: the red polygon indicates the experimental area in the Gulfof La Spezia.
(banging) activity. Moreover, we noticed that the ambient
noise included instantaneous (impulsive) sharp sounds, which
probably are due to snapping shrimp. Studies have shown
that the the Symmetric α-Stable (SαS) distribution efficiently
models snapping shrimp dominated noise [17].
Using the received signals as channel probes (i.e., running
the CEB-DFE in training mode only), we provide insight
about the variety of channel conditions during the sea trial.
The received SNRs varied between 0-40 dB depending on
the link range and the day time. Figure 3 shows the root
mean square (RMS) delay spread [18] and the Doppler spread
(averaged over all estimated multipaths) as a function of signal
reception time. One notes that the delay spread ranges from
few ms up to 30 ms. This bound is related to the particular
shallow water environment: sound rays launched over steep
angles experience high losses due to extensive reflections off
the sea boundaries. Also observe that the Doppler variability
between few Hz up to 20 Hz. Finally, the characteristic
exponent α ∈ (0, 2] of the SαS distribution of the ambient
noise is shown with respect to the daytime. The characteristic
exponent describes the impulsiveness of the ambient noise
and smaller α manifests more impulsive noise. For α=2/1
the SαS Probability Density Function (PDF) reduces to the
Gaussian/Cauchy PDF, respectively.
From the above discussion, we conclude that our signals
underwent significant distortions in both time and frequency
domain due to time varying multipath propagation. These dis-
tortions are exacerbated by impulsive noise, which corrupts the
signal at random short time intervals. Hence, achieving reliable
communications in such an environment was challenging.
III. BOOSTED TREES FOR ADAPTIVE MODULATION
In wireless radio, there exist numerous analytic formulas
that describe the BER as a function of the channel fading
model. For instance, the authors in [19] derive BER closed-
form expressions for Nakagami fading channels (which in-
cludes Rayleigh and Rician fading as special cases). Given
the lack of consensus on channel models that relate the BER
with waveform characteristics, this research adopts a pure data
driven approach that will aid machine learning algorithms to
accurately predict (regress) the BER. These techniques capture
the intricate connection between informative channel/noise
metrics and the BER. In this discussion, we assume that the
BER is dependent on the following channel metrics:
• the received SNR, SNRin, i.e., the ratio of total received
signal power over the noise power,
• the output SNR, SNRout, of the CEB-DFE [20]. It is
known that the higher the SNRout the better the equali-
sation of frequency selective fading, which in turn leads
to suppressing inter-chip interference,
• the channel average fade duration (AFD). This is a mea-
sure of rapid time-selective fading and is typically defined
as the number of times the received signal envelope falls
CMRE Reprint Series CMRE-PR-2019-025
3
Nov 21 Nov 23 Nov 252017
0
0.01
0.02
0.03
Sec
Delay spread
Nov 21 Nov 23 Nov 252017
0
5
10
15
20
Hz
Doppler spread
Nov 21 Nov 23 Nov 252017
1
1.2
1.4
1.6
1.8
2Characteristic exponent
Fig. 3. LACE17: measurements of RMS delay spread, Doppler spread and characteristic exponent α.
below a certain threshold [21]. Here, the AFD is defined
via the estimated channel energy,
E(t) =
∫ +∞
−∞|h(t, τ)|2 dτ, (4)
(where t denotes the absolute time and τ stands for the
multipath delay) and is equal to the duration for which
E(t) drops 3 dB below its maximal value. A channel with
high AFD translates to a channel with deep fades, which
in turn results to bursts of bit errors,
• the channel RMS delay spread (RMSDS) [18]. The RDS
dictates the amount of inter-chip interference that the
DFE needs to cope with,
• the Doppler spread (DS), which is a measure of the
channel time-variability. The DS is computed as the
average Power Spectral Density (PSD) over all multipath
components (channel taps).
We frame the problem of BER regression based on the
input parameters: signal name (implicitly includes coded
modulation and bit rate, see Table I), the SNRin (in dB), the
SNRout (in dB), the RMSDS (in seconds), the DS (in Hz), and
the AFD (in seconds). The regression output can be generally,
expressed as:
ˆBER = T (signal name, SNRin, SNRout, RMSDS, DS, AFD) ,(5)
where T () is a regression tree.
Regression trees [22] recursively bisect the input parameter
space to create binary partitions of the data, called nodes.
Within each node, the regression tree estimates the BER as
the average of the BER corresponding to the transmissions
assigned to that node, thus minimizing the mean squared
error (MSE) between the tree-predicted ˆBER and the measured
BER. To optimally bisect the input parameter space at each
iteration, the training algorithm selects the input parameter
and its associated value to maximally reduce the overall MSE
in a given training dataset. This node-splitting procedure is
repeated recursively until a desired MSE is achieved for the
tree, or until a desired maximum tree depth is reached. The
terminal nodes are called leaf nodes.
Regression trees can be used with heterogeneous datasets
composed of numerical, categorical (different modulation
schemes in our case), and ordinal inputs, and can handle miss-
ing inputs transparently [22]. These two properties make trees
one of the most versatile statistical learning methods currently
in use. Yet, like other statistical learning techniques, regression
trees can produce low-bias estimates but the estimates may be
susceptible to high-variance.
To alleviate the high-variance problem, researchers in statis-
tical learning have used ensembles of trees and achieved more
robust estimates [22]–[25] compared to single trees. In boosted
trees [26], the general learning technique AdaBoost [27] is
applied to regression trees. Many trees are trained on the
entire training data iteratively in such a way that at each
iteration the training samples and the predictions are assigned
weights adaptively, depending on the accuracy of the predicted
values. The aggregated, final prediction from boosted trees
is a combination of weighted predictions from each tree and
the result is that the overall boosted tree classifier produces
successively more accurate predictions as a function of the
number of trees:
ˆBER =N∑i=1
wiTi, (6)
where Ti is the i-th tree in the boosted ensemble and wi its
corresponding AdaBoost weight.
IV. RESULTS
The focus here is to find the implicit relationship between
the BER and the considered channel metrics via boosted tree
regression analysis. Recall that the experimental data were
received in relative high SNR, which is not realistic for covert
communications. For this reason, before decoding the data,
simulated white Gaussian noise is added to each individual
signal in proportion to its received SNR. We consider three
target SNRs: 0 dB, -5 dB and -10 dB. These target SNRs
are considered as three LPI regimes where the modem adapts
its throughput. Extension to a larger number of SNRs is
straightforward and is not considered in this work.
In what follows, we process 3704 audio signals that include
all seven DSSS schemes in a sequential manner. Initial training
of the DFE is performed based on a known transmitted PSK
sequence of length 1000. Table II summarises the BER per-
formance of each of the seven DSSS signals. As expected, the
CMRE Reprint Series CMRE-PR-2019-025
4
TABLE IIBER OF SIGNALS.
BPSK-4 BPSK-2 QPSK-2 BPSK-3 BPSK-1 QPSK-1 8PSK-1
bps 13 27 55 58 116 233 466
Number of signals 639 735 716 382 434 392 406
av. BER @ 0 dB 0.0086 0.0094 0.2998 0.0092 0.0046 0.0469 0.2359
av. BER @ -5 dB 0.0095 0.0197 0.4536 0.0106 0.0059 0.1188 0.3193
av. BER @ -10 dB 0.0596 0.1322 0.4767 0.0245 0.0296 0.2082 0.4664
Target SNR 0 dB
0.001 0.005 0.01 0.05Target BER
0
50
100
150
200
Thro
ughp
ut (b
ps)
Target SNR -5 dB
0.001 0.005 0.01 0.05Target BER
0
50
100
150
Thro
ughp
ut (b
ps)
Target SNR -10 dB
0.001 0.005 0.01 0.05Target BER
0
10
20
30
40
Thro
ughp
ut (b
ps) Adaptive
BPSK 13 bpsBPSK 27 bpsQPSK 55 bpsBPSK 58 bpsBPSK 116 bpsQPSK 233 bps8PSK 466 bps
Fig. 4. Average signal throughput for different BER and SNR thresholds.
performance of each signal is degraded as the SNR decreases
from 0 dB to -10 dB. Yet, our results show that the BER does
not monotonically decrease with monotonically increasing bit
rate. We believe that the proposed receiver has a difficulty in
harvesting the full coherent spreading gain in channels with
high time-variability (i.e., high Doppler spread). This issue
happens because equalization and de-spreading are not jointly
combined. This is the case, for instance, if one compares
BPSK-4 (spreading length 15) with BPSK-1 (spreading length
65). Nevertheless, this does not impede an adaptive strategy
to optimise the throughput since it is only dependent on the
accuracy of BER predictions.
The dataset of Table II is arranged in a table of 11110 rows
and seven columns. Every row i is populated with the input
parameters: signal namei, SNRini , SNRouti , RMSDSi, DSi,
AFDi and the output parameter BERi. Note that there are two
categorial inputs: the signal name and the SNRini(0 dB, -
5 dB and -10 dB). The considered ensemble of boosted trees
has 100 constituent trees and minimum leaf size of each tree
is five. To validate the performance of boosted trees, we use
disjoint sets for training and testing. In particular, we use 70%
of the data (7777 data points) for training and 30% (3333 data
points) for testing. The rationale is to provide enough training
data to learn the model sufficiently well, while ensuring that
the test set is sufficiently large to capture the inherent variety
of the data, allowing a thorough testing of the generalization
capability of the trained model. The 70/30 split of the dataset is
repeated 50 times and the average MSE of the BER prediction
is found to be 0.0012. This is a fairly accurate prediction
considering that the lowest BER achieved of our receiver is
close to 0.005.
Our main result is to compute the average signal throughputof the modem based on the following working scenario. We
assume that the modem has a trained model based on 70%
of the dataset. The modem uses the testing set (30% of
the data) to predict the BER of each DSSS signal. This is
possible by changing the signal name parameter with the
desired signal. The strategy is to select the DSSS scheme with
the fastest bit rate such that its predicted BER is bounded by
a BER threshold. After 50 random splits of the data, Figure 4
illustrates the average throughput for each of the considered
target SNRs. The superiority of the adaptive scheme over each
CMRE Reprint Series CMRE-PR-2019-025
5
fixed rate DSSS scheme is obvious for every SNR regime.
Note, for example, that the adaptive scheme is about 10 times
faster than 8PSK-1 (fastest but least reliable DSSS signal) and
20 times faster than BPSK-4 (slowest but most reliable DSSS
signal) for a target SNR=-10 dB.
V. CONCLUSIONS
The results presented in this work aim towards developing a
smart modem that will be able to boost its spectral efficiency
(bps/Hz) in response to any channel conditions while retaining
a desired level of covertness and reliability. To this end, we
investigated boosted trees as a BER predictor based on field
data. 3704 DSSS signals of different modulation orders and
baud rates were transmitted during the LACE17 trials and
our analysis showed that the boosted trees are efficient BER
predictors. In addition, the boosted trees guided our adaptation
strategy that yielded 10-20 times faster communications as
compared to a modem with a fixed rate transmission. A critical
extension of this work is to understand the impact of feedback
delay on the throughput performance. This challenge will bring
further insights on the robust application of boosted trees in
real life scenarios.
ACKNOWLEDGMENT
The authors would like to thank Joao Alves, Roberto Petroc-
cia for stimulating discussions and Piero Guerrini, Giovanni
Zappa for helping with the acquisition of the LACE17 data.
This work was supported by the NATO Allied Command
Transformation (ACT) Future Solutions Branch under the
Autonomous Security Network Programme and the Office of
Naval Research Global under grant no N62909-17-1-2093.
REFERENCES
[1] J. Proakis, Digital Communications, 4th ed., ser. Electrical engineeringseries. McGraw-Hill, 2001.
[2] T. C. Yang and W.-B. Yang, “Performance analysis of direct-sequencespread-spectrum underwater acoustic communications with low signal-to-noise-ratio input signals,” The Journal of the Acoustical Society ofAmerica, vol. 123, no. 2, pp. 842–855, Feb. 2008.
[3] T. C. Yang and W. B. Yang, “Low probability of detection underwateracoustic communications using direct-sequence spread spectrum,” TheJournal of the Acoustical Society of America, vol. 124, no. 6, pp. 3632–3647, Dec. 2008.
[4] J. Ling, H. He, J. Li, W. Roberts, and P. Stoica, “Covert underwateracoustic communications,” The Journal of the Acoustical Society ofAmerica, vol. 128, no. 5, pp. 2898–2909, Nov. 2010.
[5] B. Sherlock, J. A. Neasham, and C. C. Tsimenidis, “Spread-SpectrumTechniques for Bio-Friendly Underwater Acoustic Communications,”IEEE Access, vol. 6, pp. 4506–4520, 2018.
[9] R. C. Daniels, C. M. Caramanis, and R. W. Heath, “Adaptation inConvolutionally Coded MIMO-OFDM Wireless Systems Through Su-pervised Learning and SNR Ordering,” IEEE Transactions on VehicularTechnology, vol. 59, no. 1, pp. 114–126, Jan. 2010.
[6] X. Shu, X. Shu, J. Wang, J. Wang, H. Wang, H. Wang, X. Yang,and X. Yang, “Chaotic direct sequence spread spectrum for secureunderwater acoustic communication,” Applied Acoustics, vol. 104, pp.57–66, 2016.
[7] P. A. van Walree and G. Leus, “Robust Underwater Telemetry WithAdaptive Turbo Multiband Equalization,” IEEE Journal of OceanicEngineering, vol. 34, no. 4, pp. 645–655, Oct. 2009.
[8] H. Yigit and A. Kavak, “Adaptation using neural network in frequencyselective MIMO-OFDM systems,” in Proc. IEEE 5th InternationalSymposium on Wireless Pervasive Computing (ISWPC), Modena, Italy,May 2010, pp. 390–394.
[10] R. Daniels and R. W. Heath, “Online adaptive modulation and codingwith support vector machines,” in Proc. European Wireless Conference(EW), Lucca, Italy, Apr. 2010, pp. 718–724.
[11] J. He, H. Liu, P. Cui, J. Landon, O. Altintas, R. Vuyyuru, D. Rajan, andJ. Camp, “Design and experimental evaluation of context-aware link-level adaptation,” in Proc. IEEE INFOCOM 2012, Orlando, USA, Mar2012, pp. 2726–2730.
[12] S. Wahls and H. V. Poor, “Link Adaptation For BICM-OFDM ThroughAdaptive Kernel Regression,” in Proc. IEEE International Conference onAcoustics, Speech and Signal Processing (ICASSP), Vancouver, Canada,May 2013, pp. 5136–5140.
[13] S. Shankar, M. Chitre, and M. Jayasuria, “Data driven algorithms totune physical layer parameters of an underwater communication link,”in Proc. IEEE OCEANS 2010, Sydney, May 2010, pp. 1–5.
[14] S. Shankar and M. Chitre, “Tuning an Underwater CommunicationLink,” in Proc. MTS/IEEE OCEANS 2013, Bergen, Jun. 2013, pp. 1–9.
[15] K. Pelekanakis, L. Cazzanti, G. Zappa, and J. Alves, “Decision tree-based adaptive modulation for underwater acoustic communications,”in Proc. IEEE Third Underwater Communications and NetworkingConference (UComms), Lerici, Italy, Aug. 2016, pp. 1–5.
[16] K. Pelekanakis and M. Chitre, “Low-complexity Subband Equalizationof Mobile Underwater Acoustic Channels,” in Proc. IEEE OCEANS2015, Genova, Italy, May 2015, pp. 1–8.
[17] M. Chitre, “A high-frequency warm shallow water acoustic communica-tions channel model and measurements,” The Journal of the AcousticalSociety of America, vol. 122, no. 5, pp. 2580–2586, Nov. 2007.
[18] A. F. Molisch, “Statistical properties of the rms delay-spread of mobileradio channels with independent rayleigh-fading paths,” IEEE Transac-tions on Vehicular Technology, vol. 45, no. 1, pp. 201–204, Feb. 1996.
[19] A. Martinez, A. G. i Fabregas, and G. Caire, “A Closed-Form Ap-proximation for the Error Probability of BPSK Fading Channels,” IEEETransactions on Wireless Communications, vol. 6, no. 6, pp. 2051–2054,June 2007.
[20] M. Stojanovic, J. A. Catipovic, and J. G. Proakis, “Phase-coherent digitalcommunications for underwater acoustic channels,” IEEE Journal ofOceanic Engineering, vol. 19, no. 1, pp. 100–111, Jan 1994.
[21] N. Blaunstein and J. B. Andersen, Multipath Phenomena in CellularNetworks. Norwood, MA, USA: Artech House, Inc., 2002.
[22] T. Hastie, R.Tibshirani, and J. Friedman, The Elements of StatisticalLearning. Springer, 2001.
[23] L. Breiman, “Bagging predictors,” Machine Learning, vol. 24, no. 2, pp.123–140, 1996.
[24] ——, “Random forests,” Machine Learning, vol. 45, no. 1, pp. 5–32,2001.
[25] ——, “Arcing classifiers,” The Annals of Statistics, vol. 26, no. 3, pp.801–849, 1998.
[26] Y. Freund and R. E. Schapire, “A decision-theoretic generalization ofon-line learning and an application to boosting,” Journal of Computerand System Sciences, no. 55, pp. 119–139, 1997.
[27] R. E. Schapire, “The strength of weak learnability,” Machine Learning,no. 5, pp. 197–227, 1990.
CMRE Reprint Series CMRE-PR-2019-025
6
0B0B0BDocument Data Sheet Security Classification Project No.
Document Serial No.
CMRE-PR-2019-025
Date of Issue
May 2019
Total Pages
6 pp.
Author(s)
Konstantinos Pelekanakis, Luca Cazzanti
Title
On adaptive modulation for low SNR underwater acoustic communications
Abstract
This paper deals with adaptive underwater acoustic (UWA) communications where the receiver must operate at low signal-to-noise ratios (SNRs). The proposed modem is equipped with a set of direct sequence spread spectrum (DSSS) signals of various coding rates and modulation orders. A channel-estimate based decision feedback equalizer (CEB-DFE) is used at the receiver. We address the challenge of achieving high spectral efficiencies subject to a combination of bit-error rate (BER) and SNR constraints. To this end, adaptive selection of signals is achieved based on their BER prediction via boosted trees. This ensemble of trees learns directly from the received data and relates the BER with signal characteristics and channel metrics. The efficiency of the boosted trees is validated by post-processing thousands of acoustic signals recorded in the Gulf of La Spezia, Italy. 10-20 times faster communications as compared to a modem with a fixed rate transmission is achieved.
Keywords
Adaptive Modulation and Coding (AMC), machine learning, regression trees, Low Probability of Intercept (LPI)
Issuing Organization
NATO Science and Technology Organization Centre for Maritime Research and Experimentation
Viale San Bartolomeo 400, 19126 La Spezia, Italy
[From N. America: STO CMRE Unit 31318, Box 19, APO AE 09613-1318]
Tel: +39 0187 527 361 Fax:+39 0187 527 700
E-mail: [email protected]
